Approximate Continuous Attractor Geometry

• Approximate continuous attractor geometry is defined as the study of high-dimensional invariant sets that shape emergent behavior in nonlinear dynamical systems.
• Numerical methods such as convex optimization, sum-of-squares programming, and low-dimensional ansätze provide practical tools for reconstructing and evaluating attractor structures.
• Topological and geometric constraints, including manifold dimensionality and hyperbolicity, critically influence stability and simulation accuracy in neural and chaotic models.

Approximate continuous attractor geometry concerns the quantitative and qualitative description of high-dimensional invariant sets that govern the emergent behavior of nonlinear dynamical systems, particularly in contexts where the attractor is structured but only weakly or approximately stable. Such attractors arise in models of analog memory in neural systems, in canonical chaotic systems (e.g., Lorenz, Plykin), and in the effective description and simulation of PDEs and delay equations. Approximate geometric characterization enables both theoretical understanding and practical numerical evaluation of attractor-supported dynamics.

1. Theoretical Foundations: Manifolds, Dimensions, and Persistence

A continuous attractor is a compact, finite-dimensional, smooth manifold $M_0 = \{x_0(\theta):\theta\in\Theta\}$ invariant under the flow $\dot{x}=f_0(x)$, with marginal stability along $M_0$ and strict contraction transverse to it. Perturbations—either explicit, due to the addition of $\epsilon f_1(x)$, or implicit, due to finite-size fluctuations and noise—break this degeneracy, yielding a family of approximate slow manifolds $M_\epsilon$ that are $O(\epsilon)$-close to $M_0$ in Hausdorff norm and preserve key geometric properties for timescales $\sim \epsilon^{-1}$ (Ságodi et al., 2024). This persistent-manifold structure is guaranteed by Fenichel theory, which applies whenever $M_0$ is normally hyperbolic.

Dimensionality plays a central role: integer-valued topological dimension determines embedding and reconstructions, while fractal (box-counting, correlation) dimensions quantify the complexity of strange attractors, e.g., $D_{KY}\approx1.84$ for the Plykin attractor (Kuznetsov et al., 2019). The slow manifold inherits the symmetry group, curvature, and connectivity of the underlying task manifold—often circles ($\dot{x}=f_0(x)$0), tori, or more complicated surfaces constructed as function spaces or via embeddings.

2. Low-Dimensional Parametrizations and Neural Bump Attractors

In continuous attractor networks (CANs), geometry is operationally defined via low-dimensional functional ansätze imposed on population activity. Seeholzer et al. employ a four-parameter generalized Gaussian ansatz $\dot{x}=f_0(x)$1 on a ring manifold, with $\dot{x}=f_0(x)$2, capturing the essential translation-invariant "bump" shape in rate and spiking models (Seeholzer et al., 2017). Optimization seeks parameter values that null residuals between the ansatz and the self-consistency nonlinearity, efficiently mapping network synaptic structure to geometric features like amplitude and width.

In trained RNNs for analog memory tasks, empirical attractors manifest as one-dimensional closed curves in state space, confirmed by eigenvalue spectra consistent with theoretical slow manifolds, and geometric fidelity is upheld by delay-embedding or direct principal-plane projections (Ságodi et al., 2024).

3. Embedding and Reconstruction of Attractor Geometry

Reconstruction from partial observations is achieved via delay-coordinate embedding (Takens' theorem), where a scalar time series $\dot{x}=f_0(x)$3 reconstructs a diffeomorphic image of the $\dot{x}=f_0(x)$4-dimensional attractor in $\dot{x}=f_0(x)$5, provided $\dot{x}=f_0(x)$6 (Tajima et al., 2017). Box-counting and correlation-dimension approaches estimate the effective dimension from finite samples.

For infinite-dimensional or partially observed systems (e.g., PDEs, DDEs), embedding via finite observation maps $\dot{x}=f_0(x)$7 followed by set-oriented subdivision and continuation methods constructs finite coverings or box samples of the attractor (Gerlach et al., 2019). Diffusion maps are then applied to these sampled sets, yielding intrinsic coordinates aligned with the Riemannian geometry of the underlying invariant set, and retaining the Laplace–Beltrami structure when the normalization exponent $\dot{x}=f_0(x)$8 is used.

4. Convex Outer Approximations via Semidefinite and Sum-of-Squares Programming

Outer approximations to global attractors can be computed for polynomial systems using hierarchies of semidefinite programs (SDPs) derived from infinite-dimensional LP relaxations (Schlosser et al., 2020, Schlosser, 2022). These relaxations admit convergent positively invariant sets, often defined as sublevel sets of polynomial "almost Lyapunov" functions $\dot{x}=f_0(x)$9. The SDPs enforce convex constraints via sum-of-squares (SOS) certificates, controlling both invariance and approximation error:

$M_0$0

with volume error $M_0$1 as the polynomial degree $M_0$2 (Schlosser, 2022). Numerical examples—e.g., Van der Pol, Lorenz flows—demonstrate the method's capacity to capture both the location and fine structure of invariant sets.

5. Topological and Geometric Constraints in CANs

In biologically realistic or artificial networks, topology directly constrains the attainable geometry of continuous attractors. For instance, genuine attractor-based successor transitions in CANs trained on circular (ring) versus folded (snake) manifolds reveal irreducible geometric limitations. On the ring, all transitions have local recurrent support, while on the snake, the discontinuity at the fold precludes locally recurrent bridging; long-horizon attractor maintenance fails exactly at such topological bottlenecks, imposing a strict upper bound on transition accuracy (maximal accuracy $M_0$3 for $M_0$4 discrete points, due to the single unbridgeable gap) (Brownell, 20 Jan 2026). This demonstrates that not all manifolds are equally realizable as robust continuous attractors under given locality and stability constraints.

6. Model Reductions and One-Dimensional Approximations

Certain high-dimensional attractors admit reduction to lower-dimensional maps that encode essential dynamics. For the Lorenz attractor, an explicit reduction to a piecewise-linear Lorenz map $M_0$5 with slope $M_0$6 captures the core bifurcation structure, symbolic dynamics, and stretching/folding properties (Cholewa et al., 4 Nov 2025). The regime of validity is parameterized via the expansion rate $M_0$7, and bifurcation phenomena (period-doubling, crises, homoclinic bifurcations) are analytically linked to critical points and kneading invariants of $M_0$8.

7. Hyperbolic, Fractal, and Invariant Structure

Certain continuous attractors, such as the Plykin attractor, exemplify uniformly hyperbolic geometry, verified by explicit computation of Jacobians, Lyapunov spectra, and subspace angle distributions (Kuznetsov et al., 2019). The resulting invariant set displays band-in-band fractal structure, characterized by the Kaplan–Yorke dimension $M_0$9, and parameter charts systematically locate regions of structural stability versus bifurcation to alternate attractor types. This analytic tractability is rare, as most continuous attractors are non-hyperbolic and only possess approximate, rather than exact, invariant geometries.

---

In summary, the geometry of approximate continuous attractors is governed by an overview of topological, metric, and algebraic properties: embedding and dimension theory set bounds on observability and reconstruction; low-dimensional ansätze parameterize bump-like manifolds in neural circuits; convex optimization methods give practical and provably convergent outer approximations; and topological features set inescapable constraints on attainable attractor-supported computation. These methodologies provide a toolkit for both theoretical analysis and empirical identification of geometric structure in complex dynamical systems.